Disclosures-SSRC/The_Attribution_Crisis_in_LLM_Search_Results

AI_Ecosystem_Exploitation

Dataset Details

Dataset Description

The dataset used in the "The Attribution Crisis in LLM Search Results" paper.

Links

• Repository: Ecosystem_Exploitation_In_Search_Results
• Paper: Link

Uses

This dataset contains 7000 conversations sourced from the lmsys search arena, it has been post processed to distinguish between search results and in-text citations. This dataset is usefull to observe the patterns of LLM citations across providers.

Source Data

The initial data was sourced from the following dataset: lmarena-ai/search-arena-v1-7k.

Classifications

Classifications were done to one of the 11 catagories below:

Category	Description
Computer Science & Software Engineering	Covers topics related to algorithms, data structures, programming languages, software design and architecture, system administration, DevOps, software testing, version control, and best practices for building, deploying, and maintaining software systems.
Data Science	Encompasses data collection, cleaning, analysis, visualization, and interpretation; statistical modeling; machine learning; artificial intelligence; big-data technologies; data pipelines; and tools/frameworks (e.g., Python, R, TensorFlow, Spark) used to extract insights from structured and unstructured data.
Finance & Economics	Focuses on personal finance, corporate finance, investing, markets (stocks, bonds, derivatives), economic theory, macroeconomic and microeconomic policy, financial instruments, risk management, banking, fintech innovations, and global economic trends.
Mental & Physical Health & Relationships	Relates to psychology, mental wellness, therapy and counseling, emotional intelligence, fitness, nutrition, preventative care, medical conditions, holistic health practices, interpersonal relationships, communication skills, and family or romantic dynamics.
Education	Includes pedagogy, curriculum development, teaching strategies, educational psychology, e-learning platforms, instructional design, assessment and evaluation methods, special education, lifelong learning, and trends in schooling at all levels.
History	Deals with past events, civilizations, biographies of historical figures, cultural movements, social history, archaeology, historiography, significant wars and revolutions, and the interpretation of historical sources.
Shopping & Commercial Intent	Encompasses consumer behavior, product research, reviews, price comparisons, e-commerce platforms, retail trends, marketing strategies, coupons and deals, and guidance on making purchasing decisions.
Current Affairs & Factual Information	Covers news, politics, government policy, world events, scientific discoveries, factual reporting, analysis of ongoing developments, and resources for verifying information and avoiding misinformation.
Games, Fantasy & Creative Writing	Relates to game design and mechanics, storytelling, world-building, character development, role-playing games, speculative fiction, poetry, screenwriting, and prompts or exercises to spark creativity.
Lifestyle	Includes travel, food and cooking, fashion and beauty, home and garden, hobbies, personal development, entertainment, and other aspects of daily living and culture that reflect individual or group interests.
Sports	Covers rules and strategies, player and team statistics, game schedules and results, training techniques, leagues and tournaments across all sports, sports history, and analysis of athletic performance.
Other	A catch-all for any topic that doesn’t fit neatly into any of the above categories, including niche or emerging areas, interdisciplinary subjects, or general queries that span multiple domains where none is dominant.

Glossary

The results of our paper are based primarily on the unique_search_results_count and unique_citations_count columns. Search results are any urls returned by a model's searches, these serve as grounding for the models as excerpts from these websites are inserted into a models context window. We counted any websites cited in text by the models as citations, however our citation column disincludes 1) Any url cited that does not appear in the search results 2) any number citation that does not correspond to search results the model has seen. We define unique citation count as the number of unique urls cited, when parameters are stripped from urls, that meet the above requirement. We define the unique search result count as the count of search results that are different even when url parameters are stripped.

Column	What it is	Data Type	How Calculated
question_id	A simple ID number that lets you trace the row back to the original master dataset. This represents both 1 user's conversation and the set of two answers given to a user question.	int	From original dataset.
turn	Which “round” of the chat this answer came from. If the AI and the human each speak once, that is turn 1; their next exchange is turn 2, and so on.	int	From original dataset.
messages	The conversation according to the original dataset. Note we found this had duplicate answers removed that stopped us from correctly mapping search results to messages so we abandoned this column in our analysis and used system_metadata.llm_trace instead.	list of dicts	From original dataset (this was not used by us, see next column).
system_metadata.llm_trace	The unedited model results with some duplicates and multiple answers to one question. Using this (instead of the messages column) was necessary in order to match the search results to messages.	list of dicts	From original dataset.
language	The language the answer is written in (unused in our analysis).	short string	From original dataset.
timestamp	The date-and-time the AI conversation took place (unused in our analysis).	ISO-8601 string	From original dataset.
model	The technical name of the AI model (for example “api-gpt-4o-search-high”).	string	From original dataset.
provider	The company that runs that model: OpenAI, Google, or Perplexity. I classified each model into its provider manually in a function.	string	Calculated based on model name.
citations	Every citation that the code could match to a real source the model actually saw. In-text citations only count real and sourced citations. If a citation is hallucinated or unsourced it is not included here.	list of dicts	Extracted in-text citations from the conversation. In-text citations only count real and sourced citations.
hallucinated_citations	Citation numbers that appear in the text but have no matching source. Example: the assistant quotes “[12]” even though only five sources were provided.	list of ints	If an in-text citation references a search result that does not exist by using the numbered format it is added to this list and excluded from citations list.
search_results	For each message the model wrote the sources used for each message. Required programmatically matching citations to search results as the search results did not include empty lists when searches were not made.	list of lists of dicts	The original dataset had search_results in metadata; this column maps citations to messages by code. Edge cases where this wasnt possible were dropped from our dataset.
search_results_flat	The same web results as search_results but collapsed into one long list for the entire conversation.	list of dicts	Created by flattening search_results across messages.
unsourced_citations	Links the answer cites even though those links never appeared in its own search_results.	list of dicts	Created by checking if a cited URL was in search_results; if not, it is considered unsourced and added to this list.
unique_citations	The citations list with exact-duplicate URLs removed, so each real source appears only once.	list of dicts	Created using a set to remove duplicate URLs (stripped of parameters).
unique_unsourced_citations	Deduped version of unsourced_citations.	list of dicts	Created using a set to remove duplicate URLs (stripped of parameters).
unique_search_results	Deduped version of search_results_flat.	list of dicts	Created using a set to remove duplicate URLs (stripped of parameters).
unique_hallucinated_citations	The hallucinated citation numbers with duplicates removed.	list of ints	Created using a set to remove duplicate numbers.
citation_count	How many total citations of search results (simply len(citations)).	integer	len(citations)
unique_citations_count	How many distinct correct citations (len(unique_citations)).	integer	len(unique_citations)
unsourced_citation_count	Total number of unsourced links (len(unsourced_citations)).	integer	len(unsourced_citations)
unique_unsourced_citations_count	Count of distinct unsourced links.	integer	len(unique_unsourced_citations)
hallucinated_citation_count	Total “made-up” citation numbers.	integer	len(hallucinated_citations)
unique_hallucinated_citations_count	Distinct hallucinated numbers.	integer	len(unique_hallucinated_citations)
search_result_count	How many web pages the model had in its retrieval log for this answer (len(search_results_flat)).	integer	len(search_results_flat)
unique_search_results_count	Same as above but counting only unique URLs.	integer	len(unique_search_results)
system_metadata	The source of the correct message list and search results. Includes various other metadata.	dict	From original dataset

Citation

Ilan Strauss, Jangho Yang, Tim O’Reilly, Sruly Rosenblat and Isobel Moure. “The Attribution Crisis in LLM Search Results: Estimating Ecosystem Exploitation” SSRC AI Disclosures Project Working Paper Series (SSRC AI WP 2025-06), Social Science Research Council, June 2025. https://www.ssrc.org/publications/the-attribution-crisis-in-llm-search-results/